]> git.djapps.eu Git - pkg/ggml/sources/whisper.cpp/commit
vulkan: optimize coopmat2 dequant functions (llama/10855)
authorJeff Bolz <redacted>
Sat, 21 Dec 2024 07:04:45 +0000 (01:04 -0600)
committerGeorgi Gerganov <redacted>
Sat, 4 Jan 2025 08:45:01 +0000 (10:45 +0200)
commit39c205f55574501eb0c01073c9e16dfabd6944b6
tree6719432428e5327851188c284851126c9fd61be9
parent6d502f33dcfc93b527e5a443fbe36e83fae4a514
vulkan: optimize coopmat2 dequant functions (llama/10855)

Change the code to do 16b loads when possible and extract the appropriate
component late, so the code is effectively decoding a pair of elements and
then selecting one. This can allow more commoning to happen in the compiler
when neighboring elements are loaded.
ggml/src/ggml-vulkan/vulkan-shaders/dequant_funcs_cm2.comp